Privacy-Conscious Person Re-identification Using Low-Resolution Videos

Abstract

This paper proposes a person re-identification method for obtaining human flow information from low-resolution video generated by surveillance cameras. A requisite for the use of cameras in public spaces is protection of the privacy of individuals appearing in the captured videos. Thus, low-resolution videos (e.g. head sizes are 3-8 pixels) are expected to solve the problem of privacy, which make faces unrecognizable. However, person re-identification is more difficult in low-resolution videos than in high-resolution videos. The reason is that the person-occupied region consists of fewer pixels and has less information. Our proposed method re-identifies a person using the color features extracted from broad regions, which we consider as the most basic and important features for low-resolution videos. The color feature extraction is based on vertical relationships such as a person's head and his/her clothing because those are kept in low-resolution videos. In addition, we select the common color features, which do not change significantly between cameras. In an evaluation experiment with low-resolution videos, the re-identification accuracy of the proposed method is 71%, which is equivalent to that of manual re-identification from low-resolution videos.